Manage the context
What to include, what to cut, and why quality degrades before the window fills.
Read the guide →Diagnostic
Why the AI ignores your instructions
You told it not to do the thing. It did the thing. This is rarely defiance and almost always one of five identifiable causes.
Start by working out which failure it is
“It ignores my instructions” describes a symptom, not a cause, and the fixes are different for each cause. Before rewriting the prompt again, work out which of these you are actually looking at.
It worked, then drifted
The first few answers were fine and it went off track later. This is context dilution, not instruction failure. Your rules are still there, but they are now competing with a great deal of accumulated text.
It never followed the rule at all
Suspect placement or burial. An instruction in the middle of a long prompt gets less attention than one at the start or the end.
It follows one rule and breaks another
You probably have a conflict. Two instructions that cannot both be satisfied means one loses, and which one is not predictable.
It does exactly the thing you forbade
Negative instructions are weaker than positive ones. “Do not mention pricing” keeps pricing in the context; “only discuss delivery timelines” removes the need to.
It overdoes something you asked for once
Current models overtrigger on forceful language. If you wrote the rule in capitals with “MUST”, the emphasis itself may be the problem.
Context dilution is real and measured
The most common version of this complaint is the long conversation that slowly stops listening. That is not imagination. Independent testing across eighteen models found that performance degrades unevenly as input grows, well before the stated context limit is reached, and that degradation worsens when plausible-but-irrelevant material is present.
Two practical consequences. In a long chat, restate the constraint that matters rather than assuming the model still has it. And when a thread has drifted badly, starting a fresh conversation with a clean brief is usually faster than arguing with the old one.
Position is a lever, not a detail. Anthropic reports that placing long documents at the top of the prompt and the actual question at the end can improve response quality substantially on complex multi-document inputs. Google gives the same ordering advice. If your instructions sit above a large pasted document, move them below it.
Rewrite prohibitions as scope
The single most effective change for “it keeps doing X” is usually to stop describing what you do not want and start describing what you do. A prohibition still puts the forbidden thing into the model's working context, and it leaves the rest of the space open.
Weaker:
Do not mention our pricing. Do not speculate about the competitor. Do not be too formal. Do not use bullet points.
Stronger:
Cover only these three points: the delivery timeline, who to contact, and what we need from them.
Write in the register of a colleague who is slightly apologetic.
Write in flowing paragraphs.
If a question falls outside those three points, say it will be answered separately rather than answering it.Notice that the stronger version also tells the model what to do when the boundary is reached. A rule with no defined fallback tends to get resolved in whatever way seems most helpful, which is exactly the behaviour that felt like disobedience.
When the rule really must hold
If an instruction is genuinely non-negotiable, stop relying on the prompt for it. Check the output in code, or use a structured format so the field simply cannot contain the wrong thing. Prompt instructions are a strong preference, never a guarantee, and treating them as a guarantee is how brittle systems get built. The same reasoning applies to security, where the boundary has to be enforced outside the model entirely.
Check it properly
Turn “it seems better now” into something you can actually measure.
Read the guide →When content is hostile
Why instructions inside a document can override yours, and what a real boundary looks like.
Read the guide →Sources
- Chroma: context rot — independent testing across 18 models, July 2025
- Anthropic: prompting best practices — document ordering, and overtriggering on forceful language
- OpenAI: prompt guidance — state each instruction once
- Google: prompting strategies — put documents first and the question last